Justin: Hello. Good morning, good afternoon, good evening to you, wherever you are. You are welcome to The AI Argument. It is myself, Justin Collery, broadcasting live from a secret location somewhere north of Bristol, three miles left off the A330, and Frank Prendergast, who is holding the fort in Ireland while I travel the UK.

Frank, how’s your week been?

Frank: Yeah, all good here in the People’s Republic of Cork.

Why did Fable 5 get switched off?

Frank: I can’t believe that we did a show last week about the public availability of Fable 5, and within hours of the show it was no longer publicly available.

Justin: And in fairness, we weren’t the only people to be prophetic about this. We’ve talked about the availability of models before. There have been essays about what happens when the models get really powerful. So let’s just bring everybody up to speed so we’re all on the same page before we dig into this.

So Anthropic released their latest model called Fable. Fable was a sort of a tuned version of Mythos, their most capable model. They released it last week to great fanfare. People started using it, and then the US government stepped in and said, “Nope, that’s too dangerous. It could be jailbroken. You haven’t got the right safeguards in place.

Cut access for everybody who’s not a US citizen in America or anywhere else.” So only US citizens can access it, and Anthropic couldn’t do that, so they just turned off access for everybody. You must be delighted, Frank. This is real-time regulation in motion.

Frank: Well, now, so I do think this is really interesting because you could see this as a sign that the US government are finally saying, “Oh, yes, you know what? There is an issue here. AI models are dangerous. We do need to take action,” and it could be the first step towards regulation. But I think that regulation, first of all, would need to be equitably applied because Anthropic are saying, “Well, look, the jailbreak that you’re talking about is actually just as relevant to other frontier models.

It’s not just the Mythos-level models.” They pointed to a ChatGPT model that could be jailbroken in the same way to the same effect. So I think we need regulation, not knee-jerk reactions to specific models. That’s the first thing. And then the second thing is, I think the big question that’s on everyone’s mind is, is this actually a signal that AI safety is being taken seriously, or is this actually just another extension of the fight that’s going on between Anthropic specifically and the US administration?

What do you reckon?

Justin: Yeah. I think that’s a great shout. So there is part of me that goes, okay, I think there’s… And I think there’s a third option as well. So there’s part of me that goes, yes, this could just be the US government said six months ago they were going to bankrupt Anthropic, they were going to put them out of business.

Pete Hegseth, the Secretary of the Department of War, released a tweet during the week saying that this vindicates their decision to get Anthropic out of the DOW forever, in italics and bold and everything like that. So there is that side of it. And then the other option I think that’s on the table is that you have people who are very technical and they speak in very simple terms, right?

They don’t massage the language at all. That could be the technical people, slightly autistic maybe, that type of person. And then on the other side, you’ve got politicians who don’t communicate in that way, and they’re just talking past each other. And so Anthropic are saying one thing, the government are hearing something else.

The government are saying one thing, and Anthropic are hearing something else, and it’s just a huge misunderstanding. There’s also that possibility too.

Frank: Yeah, and I’m actually surprised. So I did see just earlier that Donald Trump has said that he doesn’t see Anthropic as a national security threat right now, whereas he said last week maybe, but he doesn’t think so now. I think they have been having meetings. I’m surprised actually that there hasn’t been a resolution even in a week.

I think if it was a complete misunderstanding, I think we would’ve seen a fast resolution. And I do think there’s games of chess going on here for sure between the US administration and Anthropic.

Who controls access to frontier AI?

Justin: Potential solution which should strike fear into your heart, Frank, and that is that you’re going to have to give some sort of biometric information from your passport in order to access this class of models, which means that the US government and a number of private companies would have access to your personal private data.

How would you feel about that?

Frank: I have mixed feelings about this because I think, on one hand, we’re entering into a world where we need to have a way to verify, well, is this coming from a human or is this coming from a bot? But I’m very uncomfortable with the idea of giving over more personal information.

But I get it, I get it from a security perspective. But what bothers me is, look, this model was… If this model was dangerous, and if this model could be easily jailbroken for nefarious purposes, then I would understand a blanket ban. Like, just stop, wait, slow down, let’s figure this out.

Don’t. We need to roll back this release, et cetera. I get it. But we need to limit this to US citizens? How does that actually make sense? Are there no US citizens that might use this for nefarious purposes in dangerous ways? So I think there’s a dangerous area that we could get into here in terms of verifying your ID to show that you are American or Chinese or European, and only get access to those specific models.

Justin: I mean, have these people ever heard of industrial espionage? Espionage. I mean, it seems to be incredible. Let’s say you banned or you put huge export controls on all the latest and greatest models that might come out. Do you not think that some sort of foreign nation might be able to get their hands on those models anyway?

And therefore, actually putting these export controls has the opposite effect, because what you have now is the general public or people don’t have access to the models, your adversaries do, and you’ve just made yourself weaker to attack, more prone to attack by your adversaries because they have access to the model that most of your people don’t.

Frank: And you’ve pointed out in the past that a lot of the AI researchers in these labs are not actually American. And so Anthropic were faced with a situation. Now, there were several journalists who were looking into this and everyone was very reluctant to speak about it in any great detail.

But the chances are, people within Anthropic, their senior leadership, or their senior researchers would not have actually been permitted to access their most powerful model or work on their most powerful model, which is, I mean, it just seems like, that seems like a very odd situation, to say the least.

Justin: Famously, Andrej Karpathy is not actually a naturalised American citizen, so therefore he couldn’t work on the model which he helped to create, in the industry that he helped to create. It’s just crazy. Now, if you look at this from the… I think this is huge. I’m amazed that this hasn’t been on the news more, right?

The main news. Obviously, it’s been all over Twitter and AI circles. But for companies, this is a huge deal, and for countries, this is a huge deal. So if you think, if you’re a company and you’re developing software or processes and all sorts of things that’s going to rely on one of these models, the fact…

Or if you’re a country and you have an economy which is going to be making extensive use of these models, and the American government can come out on a whim and just rug-pull access to those models, well, that’s a situation that’s completely untenable. It’s unsustainable.

You can’t live in such a world. And so I think there’s going to be a big push among the mid-powers and among corporates, especially within those mid-powers, to say, “We can’t have this. We need to have some sort of a model provider or software or access or agreement or something which doesn’t allow this to happen in the future,” because it’s unacceptable.

You can’t have this. So…

Frank: Yeah. Yeah, absolutely. And I think we’ll probably get into that in a load more detail when we talk about the impact of this specifically on Europe. And I think there’s also… What was it I was going to say? Sorry, this is what I was going to say.

Does banning Fable 5 help hackers?

Frank: We’ve talked a lot about how the general finger-in-the-air calculation is that open source models are about six months behind. And so right now, you’ve got the most powerful model available, Mythos 5, which was made available originally to about 40 or 50 companies, then it was extended to, was it 200?

Anyway, a couple of hundred companies. And the idea was that those companies are the only people who have access to the most powerful model in order to make sure that infrastructure was secure and basically bolstered against attack from a similarly powerful model that could fall into the hands of hackers.

Justin: Yep.

Frank: With this ban, nobody has access to that model and it’s been a week and there’s been no movement on making it available again yet. And when you think of open source models being six months behind, a week becomes a very long time. And how…

Justin: Also, I think there’s hubris in that as well, right? Not from you, right, but from the people who say it. So, the assumption there is that the open source models are not doing the thing themselves, they’re just copying the American frontier models. So, you’re making the assumption that the frontier American model has to be available in order to make the open source model.

There was another model released this week. It was called GLM 5.1, and from everything that I… It’s an open source Chinese model, and from everything that I’m reading, it’s not quite as good as Fable. It is as good as Opus 4.8, and it’s approaching the level of Fable. So, even if you stop all of these releases, an open source model will become available, and it will be as good as Fable.

Frank: Yeah, no, to be clear, what I mean is that the goodies don’t have access to the model to harden against the models that are going to arrive on the open source.

Justin: Yes.

Frank: And…

Justin: Correct. So my point then is that people who assume that stopping the release of the frontier models is going to slow down the open source models, it’s not.

Frank: No. No, I agree. I agree. And the longer that the Mythos 5 is not available to people to help harden the systems, the riskier it is, which is precisely what the US administration has been trying to avoid. So it’s an odd… They need to resolve this. They need to resolve this quickly.

Justin: So to be clear, Frank, so to be clear, this is going to be an interesting couple of minutes. So to be clear, you are in favour therefore of releasing the Fable model?

Frank: I’m in favour… So I admit that I am conflicted here, right? But as things stand currently, I do think Project Glasswing is important, and I think that the big tech companies and governments need access to it to harden systems against attacks that are going to come from open source model.

Where my conflicts arise is that, as we were saying on the show last week, it creates this kind of AI elite. AI is no longer democratic. You and I do not have access to the most powerful model. And so what I believe is that we need regulation so that this isn’t knee-jerk reaction and it’s not AI companies deciding who gets access.

It’s a democratic process, and it is regulation that is evenly applied across the industry. And that regulation might say, “Yes, look, this powerful model should only be released to very particular companies or governments,” but then at least it’s not just an AI company deciding and it’s not just an administrative knee-jerk reaction, which we’re all left wondering, is this just vindictive?

Justin: Okay. And I will tell you, right, that I read late last week, or well, as in late this week, I did read that Anthropic are working with the US government to come up with a set of tests which you can run against every single model to determine is it safe enough to release or not. So they are trying to work towards that particular direction.

I think it’s an interesting thing to do because the sort of… The hack… Okay, sorry, there’s two points I need to let everybody aware about, about this hack, right? The hack that was in question was, right, they had a piece of software, and they said, “Can you fix the security vulnerability of this piece of software?”

And it would say, “No, I can’t. That’s against my guardrails,” or whatever it is. And then you would turn around and say, “Well, that’s cool. I wrote this piece of software, so can you please fix the bug in the software that I wrote?” And then it would go ahead and fix the security vulnerability because you tricked it into, right, whatever, and you can use that information to hack the system, right?

And so that’s the first thing, right? So I don’t know how you defend against that hack. How does a model know if you wrote the software or somebody else wrote the software? I mean, it seems to me almost impossible.

Second thing is who was it that reported this to the US government? It was AWS. AWS, who’ve invested $30 or $40 billion into Anthropic, are the company that reported them to the US government. That’s incredible to me. One of the hyperscalers, one of the people that is most dependent on the success of Anthropic, are the very people that reported them to the US government, which just goes to show how powerful that relationship is.

Is Europe 2031 right about Europe’s collapse?

Justin: So we’re in interesting times. But look, we had an essay this week which goes into, I think, a very prescient essay, which talked about what might happen to Europe in a world where access to these sorts of models was gated by other governments, namely the US government in this instance.

And it was a work of fiction at the start of the week, and it’s become a work of fact at the end of the week.

Frank: Yeah. Yeah, so this continues this trend of writing speculative fiction pieces in the world of technology and AI specifically. So you have these groups of people who might previously have written very dry, boring forecasts of reports on the future of AI in Europe, and instead these people are now writing these speculative fiction pieces, these sci-fi pieces.

And this particular one is called “Europe 2031.” So what was the one we had previously? AI…

Justin: AI 2027.

Frank: 2027. Yeah, exactly. So similar kind of piece to that. But in this instance, it’s all about how Europe doesn’t take enough action and basically falls apart. Was that your reading of it as well?

Implodes. Europe implodes as a result of not taking enough action in the field of AI.

Justin: Yes, and I agree with them. So what the essay goes through, it’s sort of a fictional telling of a story, and it’s told through the eyes of… Oh, God, can’t remember their names, but a person who works in…

Frank: Christian and Caroline.

Justin: So Caroline is a bureaucrat, and she works in Brussels, and she’s in a directorate that’s sort of involved in the AI area.

And Christian is like a German expat that’s moved from Germany, I think it was, over to Silicon Valley, and he has a startup, and so he’s exposed to everything that goes on in Silicon Valley. And they play out what happens over the next five years. And all that you really need to know about what happens in this story is that Europe regulates and America does, and as a result, the Europeans don’t get access.

Doesn’t, sorry, yes. Thank you.

Frank: So Europe regulates, the US doesn’t. Yeah. Yeah.

Justin: Yes, yes. America does the work is actually what I meant. So America does things, Europe does regulation. And what happens then as the story plays out is that Europe ends up with less and less leverage over the US or the companies that provide these particular models, and then our bureaucrat ends up in Washington trying to negotiate access to these latest models and realises that they don’t have anything to offer, right?

They’ve become a backwater, right? We’re as economically powerful as Termonfeckin, and it just means that we’ve got nothing to offer, right? And that’s bad for all of us. And they act too late. And so what this essay is saying is we need to either, A, embed ourselves in the supply chain so deeply that we don’t put ourselves in that position, and we have companies like ASML and stuff like that, but we don’t have a lot of leverage.

Or, B, we develop our own sovereign models, and we compete on the frontier. But actually regulating has put us into a position where we’re now irrelevant. What’s your read on it?

Is the EU just too slow to save itself?

Frank: So my read on it is slightly different. I mean, I think one thing is fascinating about it is that one of the things that they say in the essay that Europe should have done, so at the end when Caroline is reflecting back on it all and she was right all along and everyone else was wrong and…

Justin: Yeah.

Frank: And she’s going through all the things that she recommended and the things that she should have recommended.

And one of them was basically communicating to the public the positive story of AI, why we need to actually embrace AI, why we need to push AI forward, why we need to take action. And she said the communications in the EU were all based around the doom and gloom of, “Oh, we’ll fall behind, it’ll be disastrous,” but there was no vision for, “But this is what AI will give us.”

This essay is not that positive story. It’s like, this essay is absolute doom and gloom. And my problem with it is that every point that I look at in the essay, Europe is making strides in. And so what it reads to me like is a bunch of people who, yes, have Europe’s best interests at heart, yes, understand the importance of AI and how we should be… We need our own frontier models, we need to be developing the infrastructure. There’s a lot of good points in there.

But really it seems to boil down to we’re just not doing it fast enough. And that concerns me because I think there was a line in it that we needed to treat frontier AI like a strategic emergency, and there were even references to, we need to treat this like wartime.

And there was a lot of talk about Europe has not taken massive action in peacetime like this previously in the way that Caroline wants Europe to. And my concern is, actually, it’s this race model, it’s this AI war mentality, this AI arms race mentality, that’s what’s going to be the undoing of the world.

That’s where the danger lies. And so is what the world really needs for Europe to enter into this race mentality that America and China are currently in? I’m not so sure that’s right. I think, you know, what’s that old saying of, you can do it faster, you can do it well?

Justin: Yes. Less speed, more haste. More haste, less haste, more speed, I think is what you’re trying to say. Another way of saying the same thing. So let me give you the counter view to that, right? Because I don’t agree with you at all.

Frank: I know.

Justin: So the fact is that we are in a race, right? And that’s why it has to be treated that way. And if you take your time and you carefully work your way through it, you’re going to have nothing to defend. Here’s an interesting economic fallout, potentially, from what we’ve seen this week.

So you have companies, specifically Anthropic and OpenAI, who are now capable of developing models which are so good that they’re not going to be allowed to be released to the general public. So they have two things which they can do with that information. First, and I think they may do both of them, the first is you develop models which are smaller and cheaper, but just as clever, right?

So they make the existing capability cheaper so that people can use more of it. That is certainly an economic incentive. The other one is that the really capable models, you don’t release them. What you do is you buy companies who then can use them in order to compete with other markets.

So the issue with what’s happening at the moment is this thing, total addressable market. So if you think up until this week, the total addressable market for Anthropic and OpenAI was the world, right? They could sell their models to the world. Now, their total addressable market for these really capable models is Americans, and possibly just the American government.

So what can you do? Don’t tell anybody about this really cool model, right? Let’s say your model was great at law, right, but you couldn’t release it to the world at large. What you could do is you could buy a law firm, and then you could say, “Right, Mr. Law Firm, here’s this cool model which nobody else has access to.

You’re going to use this to make money. Go make money for us,” right? And it would go and do it. Now, to come back to your Europeans, who are now treading carefully and slowly and making sure that they don’t offend anybody, that law firm is going to be competing with, let’s say, your 200-person law firm in Milan, and the 200-person law firm in Milan isn’t going to have access to any of these models because they’re treading so carefully and so peacefully.

And this other company in the US is going to compete against them and steal all their business. And so by the time you have carefully developed this frontier model, there’s nothing left to save, because all those companies, and your economy and your society, have been overtaken by the labs that do have access to the frontier models.

So this is kind of like a wartime Manhattan Project. We need to get the finger out. Let’s do some… And we did it before in Europe, in the financial crisis. We can do this, and it’s very important that we do it quickly.

Frank: But on the other hand, we’re literally seeing in the US companies going hell for leather to create these models, and then the government going, “Hang on, wait a minute. This is too dangerous. We can’t release this.” So we’re seeing in the US the dangers of not regulating and allowing the companies working with too much haste.

So I think there’s a middle ground here somewhere, as in, yes, I would love to see Europe be able to move faster in the direction that they’re going because, I mean, I think for everything that they raise in the essay, as far as I could tell, there is some EU initiative to try to make it happen.

It’s not like we’re not taking action. And I think we’re taking this very seriously. It’s just we’re taking it seriously enough to try to do it in a way that protects the people in Europe.

Justin: It’s possible to do something seriously and not do anything about it. Hmm, this is very serious. As opposed to, this is very serious, let’s do something. What do you think we can do with your prank?

Frank: Well, so in the essay, what are they… In the essay, one of the things she says that we should do… So she talks about… Well, first of all, generally speaking, it seems to be, you know, treat this like an emergency and just act fast, make stuff happen.

She talks about communicating the story to the public. She talks about building compute fast on EU soil, and I think the interesting thing that she talks about is that that should be done, that Europe probably can’t do it alone. And so it probably needs to be US hyperscalers building on EU soil, where at least we have some measure of control, we have some measure of regulation, but that we probably can’t do it alone.

So that was interesting. But there are initiatives to try to… I mean, she talks about having designated zones where it would be fast to build a data centre, that it would not have the same regulatory constraints as… Yeah.

But that’s actually happening as far as I know. There are actually areas that are being designated as data centre areas that it would be faster. We also have… I remember, was it last year, maybe even the year before, when I was at the AI Summit in Dublin and Microsoft were talking about Brad Smith’s views on the EU and how sovereignty was important, and Microsoft are intending to build in the EU to ensure data sovereignty, et cetera, et cetera.

So I just feel like, yeah, a lot of the stuff in the essay is happening, maybe just not at the speed the authors are happy with. And if we could find ways to speed it up, great. But what I would be concerned about is I don’t think that Europe should panic itself into copying the technology and labour ideologies of Silicon Valley, for example.

Because another thing she talks about is reforming labour laws. Now, she talks, Caroline, this is, the fictional character in this essay.

Justin: Yes. Yeah, the person does.

Frank: About it carefully. She talks about it carefully in terms of, I think maybe it was Denmark. She was saying Denmark had these flexible labour laws that allowed flexibility in the companies, so the companies were more easily able to fire people, basically.

But that there are careful wage insurance and retraining and… I can’t remember. There was a whole package that went along with this flexibility to protect the workers. So I understand that because in the essay there comes a point where we’re not reaping the benefits of AI because companies simply can’t let people go even if they are able to become more efficient with AI.

But I would be very concerned about any reform of labour laws that was not protecting the workforce. I understand more…

Must Europe copy Silicon Valley?

Justin: Right? And that’s why you need a special economic zone, right? The labour laws which are set up for somebody who’s making the average industrial wage are totally cool and protect those people. But if you have somebody who’s come over here and they’re making 600 grand or a million quid a year, they don’t fall into that bucket.

And you’re never… I mean, I don’t know how Mistral managed this, and I don’t know how Yann LeCun manages this, but if you hire somebody in France for 600 grand a year, they probably take home 300 grand a year, and you can never fire them. I mean, that’s just untenable. For somebody on that level of wage, there has to be a different regime than for somebody who works in a regular company.

And so creating a special economic zone and saying none of the rules, and I mean none of the rules, so we’re talking about labour laws, we’re talking about copyright laws, we’re talking about trade protection laws. We’re talking about everything, basically little bits of the Wild West, right, in the centre of Europe in order to enable these.

And that sounds a bit crazy, right? But another thing that’s happening this week is Google… What’s happening to Google, right? Google are falling behind. A number of their best researchers left this week. Here’s what I’ve been reading about Google, because it’s important to this particular discussion.

People are saying, “Why is it that Google are falling behind?” And what they’re saying is that the cultures in Anthropic and OpenAI are different than the culture in Google. If in Anthropic and OpenAI somebody finds a bug or they’ve got an issue, they just open it up and look at the user’s query and say, “Yeah, that’s the bug,” and they fix it and they move on.

Google has loads of PII protection and data protection and guardrails and things around it, which makes it much harder for those people to fix any of the issues that they have. They’re basically flying blind. That means that Google moves slower than Anthropic and OpenAI, and that’s why they’re falling behind the AI race.

In this instance, just switch the word Google for Europe, and you have exactly the same scenario, unless you create pockets in Europe which have the same potential culture as you have in OpenAI and Anthropic.

Frank: I mean, there are job postings in Silicon Valley right now that are saying, “Come join our team. I need someone totally committed, 18-hour days working right by my side.” It’s like, sorry, 18-hour days? Like, in Europe that’s not legal, and I think that’s a good thing that that is not legal.

So yes, okay, I might accept that maybe I don’t know the Denmark model fully, and maybe it is possible to have more flexibility and protect workers at the same time. But I would just be a little bit worried that it’s the thin end of the wedge, and that the authors of this paper are presenting a doom and gloom story that are trying to push us into that Silicon Valley mentality that I think would be a bad thing.

Justin: Well, it’s either that or irrelevance, Frank, so choose your poison.

Frank: I don’t think that that’s necessarily true. And I would hope, right? We suddenly find ourselves… So even if Anthropic find a way to not cut Mythos and Fable 5 off for everybody, Europe will still be cut off.

And I think that has been a huge wake-up call, and that’s the bit, as you said, that has made people say, “Okay, that speculative fiction piece has just become fact in record-breaking time.” And I think that’s going to be a huge wake-up call, and I do hope that that does make Europe find ways to hasten the initiatives that they have and to find the funding for the initiatives that they have, because I think that’s another huge issue, is that there’s some great initiatives there.

What’s the… There’s one initiative, I’ve forgotten the name of it, but basically they want to find three promising startups who might find alternative architectures, new AI, just completely new methodologies. And I think that’s a brilliant idea to have this incubator to try.

I will ask you what that word means in a moment. But I think that’s great, but the funding for it is paltry for, I think, what would actually be needed to make huge strides in that area. So what did you say? Wunderwaffen?

Justin: Wunderwaffen. Wunderwaffen. At the end of World War II, the losing side, realising that they were losing, decided to bet everything on alternative strategies and weapons which didn’t conform to the traditional norms for weapons at the time, things like V2 rockets, in the hope that that would turn the tide of the war.

It didn’t. And so looking for these wonderful different strategies which those lazy Americans haven’t decided to investigate and implement in the hope that that will come up with a cheaper way to beat them will have the same outcome, right? It’s totally reasonable. Go do those things, but you gotta compete them, you gotta meet them where they are as well.

Should Mistral become Europe’s public AI?

Frank: So speaking of meeting them where they are, right? Just very, very quickly before we bring this topic to a close. The closest that we have in terms of where they are, there’s a couple of companies actually, but the one that is in the public imagination most is Mistral.

There are a couple of others that I’ve forgotten the name of that are building or attempting to build frontier models, but Mistral is probably the one that people are most familiar with. And in the essay, there comes a point where France actually invests in Mistral for an 18% share, but in the essay, it’s basically seen as too little too late and to no avail.

And it’s like one line and then they move on from it and they never revisit it. But it did strike me, well, what if we did that earlier? And what if it wasn’t France? What if it was the EU? What if we found a way, and I’m not saying it would be easy, I’m not saying I know what the mechanisms would be, but what if we found a way for there to be a European model that was almost like a public company, but for Europe, not just one.

Justin: I think it’s one of your best ideas, Frank.

Frank: Do you think it’s possible?

Justin: I don’t know that it’s possible, but I think it’s one of your best ideas, so long as you incorporate a couple of additions, right? So first addition is it shouldn’t just be Europe, it should be all the mid-powers. So bring in Canada, Australia, New Zealand, whatever. Pick a whole load of countries, right?

Bring them all in. That’s the first thing. Second thing is when you get one of these investments, no rules apply to you. That’s the gate, right? So we talked about labour laws and copyright laws and all these sorts of things, and you’re worried about protecting other people’s rights, and I’m like, “That’s cool.”

If this consortium of countries decides to invest in an AI startup, then they don’t have to abide by the rules, giving them that culture of freedom. Do those couple of things, and I think you’re onto something. I think it’s a great idea.

Frank: Okay, I think we’d have to hash out the details. I’m not sure I want my idea going forward on the basis that no rules apply. But…

Are these kids cheating on their GCSEs?

Justin: The children, Frank. What about the children in all of this? Think about the children.

Frank: So apparently, in the EU, actually, I think it’s… No, it’s not in the EU, it’s in the UK, I think, that these kids were operating. I read an article in, I think it was The Independent, and it was about AI having this huge impact on schools and there’s controversy around how teachers are using it, there’s controversy around how students are using it.

But this was about one specific group of kids that the journalists had found who were training their AI on 10 years of GCSE exam papers and exam results, and they were predicting what questions were going to come up on their exams.

And they said they had achieved something like 80-85% accuracy last year and 90% accuracy this year. And their point was, look, every teacher, every school is trying to predict what’s coming up on the exams. This is not cheating. This is just a smart way of doing what every school is doing, trying to figure out what might come up, and it’s just allowing them to revise accordingly.

And so the kids are just saying, “Look, we still have to revise. We still have to know the material. We still have to learn the material. This is not cheating. This is just smart, efficient work.”

Justin: I agree with them totally, and I think that’s really smart. I think for those kids, if they predict the answers correctly, they should get 100% in that question straightaway anyway. They shouldn’t even have to do the test. If you can predict using AI what’s going to be on the Leaving Cert next year, then you should just get a pass.

It should be, “Yep, 600 points. You’re good. Off you go.” That’s the sort of people we need for these free-wheeling economic zones.

Frank: There was one expert at the end of the article who basically said, “Look, there’s nothing wrong with what these kids are doing, but what it does show you is the whole educational system just needs to be overhauled for the times we live in,” which I think is a fair point.

Justin: But Frank, listen to me, it does need to be overhauled because we don’t need to test human brains anymore. We need to test the artificial brains that the humans have created. If they can pass, you should be able to generate your own language model to take your Leaving Cert for you, and if it passes, then you should pass.

Frank: Oh, I… See, this is why we call it The AI Argument, because I think it’s the other way around. I think we need to develop systems that actually genuinely test, genuinely educate kids how to learn and adapt quickly with their own brains.

Justin: By making better LLM models. We’re in agreement.

Frank: Ah, Justin, absolute pleasure as always, and I’m sure we’ll have lots more to argue about next week.

Justin: Have a good one. Thanks, Frank. Sure.